The spatial relation classification task aims to categorize the spatial relations of the entities identified from text. Large Language Models (LLMs), with their outstanding language comprehension and generalization capabilities, have demonstrated strong performance across various NLP tasks. However, due to the lack of task-specific optimization for spatial relation classification, LLMs often fail to fully realize their potential in this task. In contrast, Small Language Models (SLMs) trained via supervised learning can effectively capture spatial relation patterns through task-specific training. However, their knowledge coverage and adaptability are limited compared to LLMs. In this work, we propose a hybrid framework that integrates SLMs and LLMs, leveraging Supervised In-Context Learning (ICL) to enhance spatial relation classification performance. The experimental results demonstrate that our approach outperforms the SOTA models on the SemEval 2015 dataset.